Publication thumbnail animations

The IP-SAM, TerraWeave, and EarthArena loops reveal regions of their original paper figures in the method’s reading order. Scientific labels, formulas, colors, and arrows are taken directly from the source image. A faint complete diagram provides context while the relevant regions appear. Brief mechanism highlights disappear before a clean, source-identical full-framework hold.

• IP-SAM: image encoding → foreground/background prompts → prompt-space gating → mask decoder. Purple emphasis follows the prompt mechanism.
• TerraWeave: grounded execution → verified GeoState → policy learning → skill evolution. The three original panels appear in their causal order.
• EarthArena: multi-turn protocol → evidence inputs → rubric hierarchy → scientific criteria. Blue emphasis follows evaluation and scoring.

Each directory provides an eight-second H.264 MP4 (1280 pixels wide, 24 fps) and an infinitely looping GIF (960 pixels wide, 12 fps), retaining the source figure’s aspect ratio. The clean full-frame hold runs from 6.5 to 7.65 seconds. The site swaps each loop for its original static framework in reduced-motion mode.

Edit render_publication_animations.py to change regions, timing, or highlight colors. Region coordinates use the original image pixels. Rebuild all three:

python3 assets/animations/render_publication_animations.py

Or pass --paper ip-sam, --paper terraweave, or --paper eartharena. Requirements: Python 3, Pillow, FFmpeg with libx264.